Building smarter sensors - lessons learned from computer vision

This paper describes recent work in the field of computer vision and relates the results to the much broader class of smart sensors. The sensor requirements for an autonomous mobile robot capable of map building and path planning are described. We show that when the sensor input noise is taken into consideration, a conventional CCD array is unable to provide a robust representation of an object, such that the object can be recognised regardless of the scale of the image plane. In contrast to this, biological based retinal arrays are able to achieve this (within signal limits). The paper concludes with the perspective that all sensor systems are data dependent. This is of little concern if the sensor consists of a single element, but becomes more important as larger arrays (with broader selectivity) are fabricated. These sensors may have to emulate biological systems, in an analogous manner to a retinal camera.